It is perhaps one of your prides. From an idea that takes shape in your neurons, you have the talent to transform it into a complete sentence, well-turned, in impeccable French. If your pen makes a living, in whole or in part, you may have experienced some angst the first time you used ChatGPT.
Translators know this feeling well. Since 2016, their work has been disrupted by neural machine translation systems (based on neural networks). Google Translate, Microsoft Translator Or Deep L are as formidable in translation as ChatGPT can be in writing.
As teachers specializing in different disciplines, we have worked together on projects that combine translation and journalism. As researchers for whom language is a raw material and who both use computational methods, the popularization of automated writing systems like ChatGPT challenged us.
A growing sector
Let us first look at the fears that these systems arouse. According to several experts, including Ali Zarifhonarvara doctoral student in economics from Indiana University, automated drafting systems are low-labor technologies (labour-saving technologies) that risk causing job losses in IT, communications, law and education.
However, since the arrival of artificial intelligence in translation half a dozen years ago, the job market has not dried up. It even continues to benefit from favorable prospects according to Employment Quebec.
Statistics Canada data on the labor force that worked full-year, full-time and reported employment income in the previous year shows that the number of translators, terminologists and interpreters fell from 6,270 in 2016) to 7,400 in 2021). This is an increase of 18% over five years, greater than the increase in the total for all occupations, which grew by only 6.1% during the same period.
These data highlight that the futuristic catastrophe remained a fiction. No, the profession has not been swept away by robots. Its numbers have even increased! One of the oldest professions in the world, however, had to adapt to artificial intelligence.
New challenge, new niche
Even if neural translation is now well established, the meaning of many texts still remains impenetrable for machines. This is how a new profession has emerged since 2016, which consists of revising machine translations. This operation is called the post-editingan expression that comes to us from the English meaning of publishing. Review by a human being is even part of the translation quality standardsas well as those governing the evaluation of automatic systems.
We saw how crucial human intervention remained in a study we did together ontranslation automation at the largest news agency in the country. In 2018, The Canadian Press developed Ultrad, an in-house translation system based on Google Translate. The agency’s journalists can use it to translate dispatches from their English-speaking colleagues or from the Associated Press. The table below shows some of the errors made by the system and the corrections made thanks to the vigilance of the journalists.
|Source (English)||Automatic translation (Ultrad)||Post-editing (human)|
|Steven Guilbeault will table a new greenhouse gas emissions plan in Parliament this morning.||Steven Guilbeault will table in Parliament this morning a new greenhouse gas emissions plan.||Steven Guilbeault will table a new plan reduction greenhouse gas emissions.|
|lich was arrested Feb. 17 years old initially denied lease||lich been stopped February 17 and initially denied bail||Ms Lich had been stopped February 17, in Ottawa. His bail request was initially denied.|
|The province says the more than $5-billion investment||Province claims that the investment of more than $5 billion||The Ontario government claims that the investment of more than $5 billion|
In the first, the system did not know that the emission reduction, implied in English, had to be specified in French.
In the second, he masculinized one of the leaders of the truckers’ movement who occupied downtown Ottawa in 2022. He also understood that she had refused her own release, whereas this one had in fact been rejected by a court.
In the third, he did not take into account that the employment of province figuratively to refer to a government is allowed in English, but not in French.
Some errors, however, have escaped the notice of humans. Word sectionwhich designates in English an article in an act, was erroneously translated by “section” in a text on Bill C-11 which was published in different items in March 2022.
Beware of mechanical objectivity
Whereas translators previously had to start with a text in the source language, machine translation offers them a seemingly complete and well-turned first version. This primer has all the appearance of a job well done and yielding to it induces what is called the“priming effect” (priming-effect).
This effect can also be reinforced bymechanical objectivity. Throughout history, scientists have sought to eliminate human subjectivity in their work. The use of measuring devices is therefore associated with objectivity and neutrality. A certain epistemic authority accompanies their use, authority which is also conferred on systems based on artificial intelligence.
Translators are familiar with these phenomena and have learned to be wary of them. Beneath the polished surface of the texts produced by machine translation systems and conversational robots like ChatGPT surreptitiously hide errors of various kinds that the priming effect hides from inattentive readers.
Post-editing work has accustomed translation professionals to recognizing them, as Thierry Grass, translator and professor of translation, explains in his article “To err is not human”. In particular, he tells us that automated translation systems produce texts that appear to be perfect in terms of form, but which may contain flaws in terms of substance, content and logic.
Translation professionals have been sort of scouts who can teach us how to deal with automated writing systems like ChatGPT. And this is how they paved the way for a critical use of automated writing systems and a better understanding of their limits: fallacious, equivocal reasoning, shortcuts, ellipsis, received ideas, so many new headings for the classic by Normand Baillargeon, A short course in intellectual self-defense.